Free AI SEO Audit: A Do-It-Yourself Method

An AI answer about your brand passes through four stages, and each one fails in a different way.

• Parametric. This is what the model learned in training, frozen at a knowledge cutoff. OpenAI's model documentation lists a February 16, 2026 cutoff for the GPT-5.6 family and August 31, 2025 for ChatGPT's Instant path. Nothing you publish today reaches an answer given from memory alone.
• Access. A crawler must be allowed to fetch your page and be able to reach it. robots.txt, your firewall or CDN, and JavaScript rendering decide this.
• Retrieval. Your page has to enter the model's context for a given question. Engines rewrite the user's prompt into their own search queries, so you never see the real input.
• Generation. The model picks which retrieved pages to use and cite.

Retrieval is discretionary on every platform. As of October 2026, ChatGPT decides per question whether to search, web search in Claude.ai is a user toggle, and Google AI Overviews often don't trigger at all. Published estimates of how often ChatGPT searches run from about 18% of conversations (Profound) to 80.5% of a 100-prompt set (cloro), because the two measure different things.

So any answer you see might come from live search or from training data. You can only tell when it shows citations. For your audit, record whether each answer cited sources. An uncited answer tells you about the model's memory. A cited answer tells you about access and retrieval.

Be clear about this before you start, or you'll chase things nobody can move.

In your control

• Which AI crawlers your robots.txt allows, bot by bot.
• Whether your firewall or CDN lets those crawlers through.
• Whether key facts (prices, product names, what you do) are in the HTML before JavaScript runs.
• Whether noindex or snippet controls are hiding pages by accident.
• Whether the facts on your own site agree with each other.

Partly in your control

• Being named on third-party pages that engines already cite. The evidence shows a strong association between that and appearing in answers, with no promised effect.
• Stating a specific fact plainly on a page that can be fetched. That gives an engine a sentence it can quote.

Not in your control

• What a model learned in training, this quarter.
• Which pages a retrieval system picks and in what order.
• How the model phrases the answer.

The research is blunt about the last group. A July 2026 critical survey of 45 GEO studies (arXiv 2607.14035) found that no reviewed technique shows a stable, cross-platform causal effect on discoverability. That is why this audit splits into two halves: observe what the engines say, then fix what blocks you. If you want the comparison with a classic crawl, see how an AI visibility audit differs from an SEO audit.

Pick five to ten questions a buyer would actually ask. Mix four types:

1. Category: "best scheduling software for physiotherapy clinics".
2. Comparison: "[your brand] vs [main competitor]".
3. Problem: the pain your product solves, phrased the way a customer says it.
4. Brand facts: "how much does [your brand] cost", "who owns [your brand]".

Say you run a dental clinic in Austin. A category question might be "best family dentist in south Austin open Saturdays". A brand fact question might be "does [clinic name] take new patients".

Now fix the conditions so results are comparable:

• Log out, or turn memory off. OpenAI says ChatGPT may use saved memories when rewriting a search query, so a logged-in answer reflects your history.
• Run everything on the same day.
• Use the exact same wording in every engine.
• Record the model name where the app shows it.

Keep the question list in a spreadsheet. You'll reuse it every time you re-run the audit, and the spreadsheet method for share of voice uses the same structure.

Ask each question in ChatGPT, Perplexity, Claude, the Gemini app, Google AI Overviews and Google AI Mode. Treat the three Google surfaces as separate products. Google states that AI Overviews and AI Mode "may use different models and techniques" and can return different links.

For every answer, log:

• Engine, date and question.
• Whether your brand is named.
• Which competitors are named.
• Whether citations appeared, and which URLs.
• Every fact stated about you.

Then grade each fact as right, wrong or unverifiable. Common errors to look for:

• An old or wrong price.
• A retired product name or a feature you don't offer.
• A wrong founding year, location or owner.
• Your brand confused with a competitor.

Keep "wrong" and "unverifiable" apart. A provably wrong price is a fix target. A claim you simply can't check is a different problem. If an engine errors out or refuses, log that too, and don't count it as a clean answer (here's why failed responses matter).

Finally, check your own site for the same facts. If your pricing page, about page and footer disagree, an engine has conflicting sentences to choose from. That is fully in your hands.

Access problems remove you silently. Work through these in order.

1. robots.txt, per bot. Rules are independent for each crawler. OpenAI's crawler page says GPTBot governs training, OAI-SearchBot governs ChatGPT search, and "each setting is independent of the others". This is valid and keeps you in ChatGPT search while opting out of training:

User-agent: OAI-SearchBot
Allow: /

User-agent: GPTBot
Disallow: /

Also check PerplexityBot, Claude-SearchBot and Meta-WebIndexer.

2. Firewall and CDN. OpenAI's help centre says eligibility needs both: allow OAI-SearchBot in robots.txt and confirm your host or CDN allows traffic from OpenAI's published searchbot IP addresses. A 403 to a crawler that robots.txt allows is a complete, invisible failure. From September 15, 2026, Cloudflare blocks Training and Agent crawlers by default on ad-displaying pages for new domains and free-tier customers, while Search stays allowed. A site can lose some AI access without anyone editing anything. The AI Crawler Checker requests a page as each AI crawler and shows which ones get turned away.

3. JavaScript rendering. Vercel found that ChatGPT and Claude crawlers fetch JavaScript files but don't execute them, so they can't read client-side rendered content. View your page source with JavaScript off. If prices or product names only appear after the page loads, most AI crawlers never see them. Googlebot does render, so this matters for non-Google engines.

4. Snippet controls. noindex, nosnippet, max-snippet:0 and data-nosnippet keep content out of AI answers on Google and Bing. Check templates for accidental use.

5. Index status, engine by engine. Google says a page must be indexed and eligible for a snippet to appear as a supporting link in AI Overviews or AI Mode. Check Google Search Console and Bing Webmaster Tools separately. In Otterly's 2026 experiment, Google deindexed two test sites and its AI surfaces stopped citing them, while Bing kept them and Microsoft Copilot supplied 2,718 of one site's 3,934 tracked citations.

For the full checklist, follow the AI crawler access audit playbook or the shorter free GEO audit.

Your audit captures one day. Treat it as a baseline, not a verdict.

Answers vary for reasons unrelated to your site:

• The same question can trigger search on one run and not the next.
• Models get updated, and apps route questions to different models.
• Engines rewrite prompts differently each time.
• Search results behind the answer shift.

So one "not mentioned" is a data point. Five "not mentioned" across three months is a pattern. Re-run the same questions under the same conditions on a fixed schedule, and only compare like with like. The quarterly AI visibility review gives you a cadence.

Two more limits. First, there is no rank to record. Google's AI report in Search Console shows impressions only, with no position data, so any AI "position" has nothing to be checked against (more on that). Second, don't blend engines into one number. In Otterly's test, one engine cited the sites thousands of times while four barely did. An average would have hidden that.

If you want a quick second read, the free checker on this site runs up to three buyer questions and scores the answers. It is also a snapshot.

• Treating GPTBot as the search switch. Blocking GPTBot doesn't remove you from ChatGPT search, and allowing it doesn't get you found. OAI-SearchBot controls search.
• Checking robots.txt and stopping. Robots.txt is necessary but not sufficient. Test what the firewall actually returns.
• Testing while logged in. Memory and custom instructions change the answer. Log out.
• Reading citations as a ranking. Being cited is not a position, and cited is not the same as used.
• Adding schema or an llms.txt as the fix. Ahrefs' matched test found AI Overviews citations fell 4.6% after pages added schema. Google says Search ignores llms.txt. Existing markup does no harm, so don't rip it out, but it isn't an AI lever.
• Rewriting pages in an "AI-friendly" style. Google says you don't need to write a specific way for generative AI search. C-SEO Bench found most conversational SEO methods largely ineffective and often harmful.
• Hidden text or instructions aimed at AI. Prompt injection and cloaking for AI crawlers exist. They are illegitimate, they risk removal, and they can backfire on your reputation. Don't use them.

The fix list that survives the evidence is short: open access, server-render your key facts, remove accidental suppression, and make your own site agree with itself.

How often should I re-run a DIY AI visibility audit? Quarterly works for most teams, plus once after any big change: a site migration, new pricing, a CDN or firewall change, or a rebrand. Use the exact same questions and conditions each time, logged out and on one day. Changing the question set between runs makes the results impossible to compare, so add new questions as a separate group rather than swapping old ones out.

Do I need a paid tool to audit AI visibility? No. The method on this page needs a spreadsheet, a logged-out browser and an hour or two. Paid tools mainly save time on repetition: running many questions across engines on a schedule and keeping history. Whatever you use, check that it reports each engine separately and says whether answers came with live search, because a blended number hides the differences that matter.

Why does ChatGPT give a different answer each time I ask? Model outputs vary between runs, and ChatGPT decides per question whether to search the web. One run may answer from training data and the next may pull live results with citations. ChatGPT also rewrites your prompt into its own search queries, which can differ run to run. That's why a single answer is a sample, and why you should log whether citations appeared.

Can I fix a wrong fact an AI engine states about my brand? You can fix your side. State the correct fact plainly on a page crawlers can fetch, in the HTML before JavaScript runs, and remove contradictions across your own pages. That gives search-enabled answers a correct sentence to quote. You can't edit a model's training data, so answers given from memory may stay wrong until the vendor ships a newer model.

Does blocking GPTBot remove my site from ChatGPT? No. OpenAI documents GPTBot as its training crawler and OAI-SearchBot as the crawler for ChatGPT search, and says each setting is independent. You can block GPTBot to opt out of training and still allow OAI-SearchBot. Blocking OAI-SearchBot is what keeps you out of ChatGPT search answers, though pages can still appear as navigational links.